ICCV 2023 Papers

Combined with the generative prior of pre-trained models and the flexibility of text, text-driven generative domain adaptation can generate images from a wide range of target domains. However, current methods still suffer from overfitting and the mode collapse problem. In this paper, we analyze the mode collapse from the geometric point of view and reveal its relationship to the Hessian matrix of generator. To alleviate it, we propose the spectral consistency regularization to preserve the diversity of source domain without restricting the semantic adaptation to target domain. We also design granularity adaptive regularization to flexibly control the balance between diversity and stylization for target model. We conduct experiments for broad target domains compared with state-of-the-art methods and extensive ablation studies. The experiments demonstrate the effectiveness of our method to preserve the diversity of source domain and generate high fidelity target images.

Token-based masked generative models are gaining popularity for their fast inference time with parallel decoding. While recent token-based approaches achieve competitive performance to diffusion-based models, their generation performance is still suboptimal as they sample multiple tokens simultaneously without considering the dependence among them. We empirically investigate this problem and propose a learnable sampling model, Text-Conditioned Token Selection (TCTS), to select optimal tokens via localized supervision with text information. TCTS improves not only the image quality but also the semantic alignment of the generated images with the given texts. To further improve the image quality, we introduce a cohesive sampling strategy, Frequency Adaptive Sampling (FAS), to each group of tokens divided according to the self-attention maps. We validate the efficacy of TCTS combined with FAS with various generative tasks, demonstrating that it significantly outperforms the baselines in image-text alignment and image quality. Our text-conditioned sampling framework further reduces the original inference time by more than 50% without modifying the original generative model.

Neural Radiance Fields (NeRFs) are a very recent and very popular approach for the problems of novel view synthesis and 3D reconstruction. A popular scene representation used by NeRFs is to combine a uniform, voxel-based subdivision of the scene with an MLP. Based on the observation that a (sparse) point cloud of the scene is often available, this paper proposes to use an adaptive representation based on tetrahedra obtained by Delaunay triangulation instead of uniform subdivision or point-based representations. We show that such a representation enables efficient training and leads to state-of-the-art results. Our approach elegantly combines concepts from 3D geometry processing, triangle-based rendering, and modern neural radiance fields. Compared to voxel-based representations, ours provides more detail around parts of the scene likely to be close to the surface. Compared to point-based representations, our approach achieves better performance. The source code is publicly available at: https://jkulhanek.com/tetra-nerf.

Current motion forecasting approaches typically train a deep end-to-end model from the source domain data, and then apply it directly to target subjects. Despite promising results, they remain non-optimal, due to privacy considerations, the test person and his/her natural properties (e.g., stature, behavioral trait) are typically unseen/absent in training. In this case, the source pre-trained model has a low ability to adapt to these out-of-source characteristics, resulting in an unreliable prediction. To tackle this issue, we propose a novel helper-predictor test-time personalization approach (H/P-TTP), which allows for a generalizable representation of out-of-source subjects to gain more realistic predictions. Concretely, the helper is preceded by explicit and implicit augmenters, where the former yields noisy sequences to improve robustness, while the latter is to generate novel-domain data with an adversarial learning paradigm. Then, the domain-generalizable learning is achieved where the helper can extract cross-subject invariant-knowledge to update the predictor. At test time, given a new person, the predictor is able to be further optimized to empower personalized capabilities to the specific properties. Under several benchmarks, extensive experiments show that with H/P-TTP, the existing predictive models are significantly improved for various unseen subjects.

While the design of blind image quality assessment (IQA) algorithms has improved significantly, the distribution shift between the training and testing scenarios often leads to a poor performance of these methods at inference time. This motivates the study of test time adaptation (TTA) techniques to improve their performance at inference time. Existing auxiliary tasks and loss functions used for TTA may not be relevant for quality-aware adaptation of the pre-trained model. In this work, we introduce two novel quality-relevant auxiliary tasks at the batch and sample levels to enable TTA for blind IQA. In particular, we introduce a group contrastive loss at the batch level and a relative rank loss at the sample level to make the model quality aware and adapt to the target data. Our experiments reveal that even using a small batch of images from the test distribution helps achieve significant improvement in performance by updating the batch normalization statistics of the source model.

Spiking Neural Networks (SNNs) offer a highly promising computing paradigm due to their biological plausibility, exceptional spatiotemporal information processing capability and low power consumption. As a temporal encoding scheme for SNNs, Time-To-First-Spike (TTFS) encodes information using the timing of a single spike, which allows spiking neurons to transmit information through sparse spike trains and results in lower power consumption and higher computational efficiency compared to traditional rate-based encoding counterparts. However, despite the advantages of the TTFS encoding scheme, the effective and efficient training of TTFS-based deep SNNs remains a significant and open research problem. In this work, we first examine the factors underlying the limitations of applying existing TTFS-based learning algorithms to deep SNNs. Specifically, we investigate issues related to over-sparsity of spikes and the complexity of finding the `causal set'. We then propose a simple yet efficient dynamic firing threshold (DFT) mechanism for spiking neurons to address these issues. Building upon the proposed DFT mechanism, we further introduce a novel direct training algorithm for TTFS-based deep SNNs, called DTA-TTFS. This method utilizes event-driven processing and spike timing to enable efficient learning of deep SNNs. The proposed training method was validated on the image classification task and experimental results clearly demonstrate that our proposed method achieves state-of-the-art accuracy in comparison to existing TTFS-based learning algorithms, while maintaining high levels of sparsity and energy efficiency on neuromorphic inference accelerator.

In this paper, we propose a new paradigm, named Historical Object Prediction (HoP) for multi-view 3D detection to leverage temporal information more effectively. The HoP approach is straightforward: given the current timestamp t, we generate a pseudo Bird's-Eye View (BEV) feature of timestamp t-k from its adjacent frames and utilize this feature to predict the object set at timestamp t-k. Our approach is motivated by the observation that enforcing the detector to capture both the spatial location and temporal motion of objects occurring at historical timestamps can lead to more accurate BEV feature learning. First, we elaborately design short-term and long-term temporal decoders, which can generate the pseudo BEV feature for timestamp t-k without the involvement of its corresponding camera images. Second, an additional object decoder is flexibly attached to predict the object targets using the generated pseudo BEV feature. Note that we only perform HoP during training, thus the proposed method does not introduce extra overheads during inference. As a plug-and-play approach, HoP can be easily incorporated into state-of-the-art BEV detection frameworks, including BEVFormer and BEVDet series. Furthermore, the auxiliary HoP approach is complementary to prevalent temporal modeling methods, leading to significant performance gains. Extensive experiments are conducted to evaluate the effectiveness of the proposed HoP on the nuScenes dataset. We choose the representative methods, including BEVFormer and BEVDet4D-Depth to evaluate our method. Surprisingly, HoP achieves 68.2% NDS and 61.6% mAP with ViT-L on nuScenes test, outperforming all the 3D object detectors on the leaderboard by a large margin.

Referring video object segmentation aims to segment a referent throughout a video sequence according to a natural language expression. It requires aligning the natural language expression with the objects' motions and their dynamic associations at the global video level but segmenting objects at the frame level. To achieve this goal, we propose to simultaneously maintain a global referent token and a sequence of object queries, where the former is responsible for capturing video-level referent according to the language expression, while the latter serves to better locate and segment objects with each frame. Furthermore, to explicitly capture object motions and spatial-temporal cross-modal reasoning over objects, we propose a novel temporal collection-distribution mechanism for interacting between the global referent token and object queries. Specifically, the temporal collection mechanism collects global information for the referent token from object queries to the temporal motions to the language expression. In turn, the temporal distribution first distributes the referent token to the referent sequence across all frames and then performs efficient cross-frame reasoning between the referent sequence and object queries in every frame. Experimental results show that our method outperforms state-of-the-art methods on all benchmarks consistently and significantly.

Targeting for detecting anomalies of various sizes for complicated normal patterns, we propose a Template-guided Hierarchical Feature Restoration method, which introduces two key techniques, bottleneck compression and template-guided compensation, for anomaly-free feature restoration. Specially, our framework compresses hierarchical features of an image by bottleneck structure to preserve the most crucial features shared among normal samples. We design template-guided compensation to restore the distorted features towards anomaly-free features. Particularly, we choose the most similar normal sample as the template and leverage hierarchical features from the template to compensate the distorted features. The bottleneck could partially filter out anomaly features, while the compensation further converts the reminding anomaly features towards normal with template guidance. Finally, anomalies are detected in terms of the cosine distance between the pre-trained features of an inference image and the corresponding restored anomaly-free features. Experimental results demonstrate the effectiveness of our approach, which achieves the state-of-the-art performance on the MVTec LOCO AD dataset.

Face recognition systems are increasingly being used in different applications. In such systems, some features (also known as embeddings or templates) are extracted from each face image. Then, the extracted templates are stored in the system's database during the enrollment stage and are later used for recognition. In this paper, we focus on template inversion attacks against face recognition systems and introduce a novel method (dubbed GaFaR) to reconstruct 3D face from facial templates. To this end, we use a geometry-aware generator network based on generative neural radiance fields (GNeRF), and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our network with a semi-supervised learning approach using real and synthetic images simultaneously. For the real training data, we use a Generative Adversarial Network (GAN) based framework to learn the distribution of the latent space. For the synthetic training data, where we have the true latent code, we directly train in the latent space of the generator network. In addition, during the inference stage, we also propose optimization on the camera parameters to generate face images to improve the success attack rate (up to 17.14% in our experiments). We evaluate the performance of our method in the whitebox and blackbox attacks against state-of-the-art face recognition models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates. The project page is available at: https://www.idiap.ch/paper/gafar

Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-Adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-Adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method.

Large vision-language models, such as CLIP, learn robust representations of text and images, facilitating advances in many downstream tasks, including zero-shot classification and text-to-image generation. However, these models have several well-documented limitations. They fail to encapsulate compositional concepts, such as counting. To the best of our knowledge, this work is the first to extend CLIP to handle object counting. We introduce a simple yet effective method to improve the quantitative understanding of vision-language models, while maintaining their overall performance on common benchmarks. Our method automatically augments image captions to create hard negative samples that differ from the original captions by only the number of objects. For example, an image of three dogs can be contrasted with the negative caption "Six dogs playing in the yard". A dedicated loss encourages discrimination between the correct caption and its negative variant. In addition, we introduce CountBench, a new benchmark for evaluating a model's understanding of object counting, and demonstrate significant improvement over baseline models on this task. Furthermore, we leverage our improved CLIP representations for text-conditioned image generation, and show that our model can produce specific counts of objects more reliably than existing ones.

Video anomaly detection (VAD) without human monitoring is a complex computer vision task that can have a positive impact on society if implemented successfully. While recent advances have made significant progress in solving this task, most existing approaches overlook a critical real-world concern: privacy. With the increasing popularity of artificial intelligence technologies, it becomes crucial to implement proper AI ethics into their development. Privacy leakage in VAD allows models to pick up and amplify unnecessary biases related to people's personal information, which may lead to undesirable decision making. In this paper, we propose TeD-SPAD, a privacy-aware video anomaly detection framework that destroys visual private information in a self-supervised manner. In particular, we propose the use of a temporally-distinct triplet loss to promote temporally discriminative features, which complements current weakly-supervised VAD methods. Using TeD-SPAD, we achieve a positive trade-off between privacy protection and utility anomaly detection performance on three popular weakly supervised VAD datasets: UCF-Crime, XD-Violence, and ShanghaiTech. Our proposed anonymization model reduces private attribute prediction by 32.25% while only reducing frame-level ROC AUC on the UCF-Crime anomaly detection dataset by 3.69%.

The success of automated medical image analysis depends on large-scale and expert-annotated training sets. Unsupervised domain adaptation (UDA) has been raised as a promising approach to alleviate the burden of labeled data collection. However, they generally operate under the closed-set adaptation setting assuming an identical label set between the source and target domains, which is over-restrictive in clinical practice where new classes commonly exist across datasets due to taxonomic inconsistency. While several methods have been presented to tackle both domain shifts and incoherent label sets, none of them take into account the common characteristics of the two issues and consider the learning dynamics along network training. In this work, we propose optimization trajectory distillation, a unified approach to address the two technical challenges from a new perspective. It exploits the low-rank nature of gradient space and devises a dual-stream distillation algorithm to regularize the learning dynamics of insufficiently annotated domain and classes with the external guidance obtained from reliable sources. Our approach resolves the issue of inadequate navigation along network optimization, which is the major obstacle in the taxonomy adaptive cross-domain adaptation scenario. We evaluate the proposed method extensively on several tasks towards various endpoints with clinical significance. The results demonstrate its effectiveness and improvements over previous methods.

Learning discriminative task-specific features simultaneously for multiple distinct tasks is a fundamental problem in multi-task learning. Recent state-of-the-art models consider directly decoding task-specific features from one shared task-generic feature (e.g., feature from a backbone layer), and utilize carefully designed decoders to produce multi-task features. However, as the input feature is fully shared and each task decoder also shares decoding parameters for different input samples, it leads to a static feature decoding process, producing less discriminative task-specific representations. To tackle this limitation, we propose TaskExpert, a novel multi-task mixture-of-experts model that enables learning multiple representative task-generic feature spaces and decoding task-specific features in a dynamic manner. Specifically, TaskExpert introduces a set of expert networks to decompose the backbone feature into several representative task-generic features. Then, the task-specific features are decoded by using dynamic task-specific gating networks operating on the decomposed task-generic features. Furthermore, to establish long-range modeling of the task-specific representations from different layers of TaskExpert, we design a multi-task feature memory that updates at each layer and acts as an additional feature expert for dynamic task-specific feature decoding. Extensive experiments demonstrate that our TaskExpert clearly outperforms previous best-performing methods on all 9 metrics of two competitive multi-task learning benchmarks for visual scene understanding (i.e., PASCAL-Context and NYUD-v2). Code and models will be made publicly available.

Although existing few-shot learning works yield promising results for in-domain queries, they still suffer from weak cross-domain generalization. Limited support data requires effective knowledge transfer, but domain-shift makes this harder. Towards this emerging challenge, researchers improved adaptation by introducing task-specific parameters, which are directly optimized and estimated for each task. However, adding a fixed number of additional parameters fails to consider the diverse domain shifts between target tasks and the source domain, limiting efficacy. In this paper, we first observe the dependence of task-specific parameter configuration on the target task. Abundant task-specific parameters may over-fit, and insufficient task-specific parameters may result in under-adaptation -- but the optimal task-specific configuration varies for different test tasks. Based on these findings, we propose the Task-aware Adaptive Network (TA2-Net), which is trained by reinforcement learning to adaptively estimate the optimal task-specific parameter configuration for each test task. It learns, for example, that tasks with significant domain shift usually have a larger need for task-specific parameters for adaptation. We evaluate our model on Meta-dataset. Empirical results show that our model outperforms existing state-of-the-art methods.

Prompt learning has become one of the most efficient paradigms for adapting large pre-trained vision-language models to downstream tasks. Current state-of-the-art methods, like CoOp and ProDA, tend to adopt soft prompts to learn an appropriate prompt for each specific task. Recent CoCoOp further boosts the base-to-new generalization performance via an image-conditional prompt. However, it directly fuses identical image semantics to prompts of different labels and significantly weakens the discrimination among different classes as shown in our experiments. Motivated by this observation, we first propose a class-aware text prompt (CTP) to enrich generated prompts with label-related image information. Unlike CoCoOp, CTP can effectively involve image semantics and avoid introducing extra ambiguities into different prompts. On the other hand, instead of reserving the complete image representations, we propose text-guided feature tuning (TFT) to make the image branch attend to class-related representation. A contrastive loss is employed to align such augmented text and image representations on downstream tasks. In this way, the image-to-text CTP and text-to-image TFT can be mutually promoted to enhance the adaptation of VLMs for downstream tasks. Extensive experiments demonstrate that our method outperforms the existing methods by a significant margin. Especially, compared to CoCoOp, we achieve an average improvement of 4.03% on new classes and 3.19% on harmonic-mean over eleven classification benchmarks.

In many video restoration/translation tasks, image processing operations are naively extended to the video domain by processing each frame independently, disregarding the temporal connection of the video frames. This disregard for the temporal connection often leads to severe temporal inconsistencies. State-Of-The-Art (SOTA) techniques that address these inconsistencies rely on the availability of unprocessed videos to implicitly siphon and utilize consistent video dynamics to restore the temporal consistency of frame-wise processed videos which often jeopardizes the translation effect. We propose a general framework for this task that learns to infer and utilize consistent motion dynamics from inconsistent videos to mitigate the temporal flicker while preserving the perceptual quality for both the temporally neighboring and relatively distant frames without requiring the raw videos at test time. The proposed framework produces SOTA results on two benchmark datasets, DAVIS and videvo.net, processed by numerous image processing applications. The code and the trained models will be open-sourced upon acceptance.

In this paper we introduce the Tangent Sampson error, which is a generalization of the classical Sampson error in two-view geometry that allows for arbitrary central camera models. It only requires local gradients of the distortion map at the original correspondences (allowing for pre-computation) resulting in a negligible increase in computational cost when used in RANSAC or local refinement. The error effectively approximates the true-reprojection error for a large variety of cameras, including extremely wide field-of-view lenses that cannot be undistorted to a single pinhole image. We show experimentally that the new error outperforms competing approaches both when used for model scoring in RANSAC and for non-linear refinement of the relative camera pose.